Financial transaction risk control method and device, storage medium and program product

Abstract

The embodiment of the invention provides a financial transaction risk control method and device, a storage medium and a program product, and relates to the technical field of artificial intelligence. The method comprises the steps of obtaining original data of a financial transaction, performing data preprocessing, and outputting a standardized feature vector; determining a scene probability and an enhanced feature according to the standardized feature vector and a preset scene adaptive attention mechanism; determining an interaction strength coefficient according to the enhanced feature and the scene probability; inputting the enhanced features and the interaction strength coefficient into a prediction model to obtain a risk prediction result of risk prediction; and executing a multi-level decision according to the risk prediction result and the scene probability, and determining a target output result corresponding to the original data. According to the scheme of the invention, the problem that the existing financial transaction risk prediction technology cannot effectively deal with the dynamics of a financial transaction scene, the evolution of a risk mode and the speciality of business requirements is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a financial transaction risk control method and device, a storage medium and a program product. BACKGROUND

[0002] With the rapid development of financial technology and the comprehensive popularization of digital payment, the scale of financial transactions presents an explosive growth, the transaction scene becomes increasingly complex, the concealment and variability of transaction risk are significantly improved, and transaction risk identification has become a core technical link to protect the safety of the financial system and maintain the rights and interests of user funds. At present, the mainstream financial transaction risk prediction technology solutions in the industry mainly fall into two categories, namely linear models and tree-based ensemble models, but both solutions have obvious defects in actual application, making it difficult to meet the high-precision risk identification needs in complex dynamic scenarios.

[0003] Among them, the linear model assigns fixed weights to each transaction feature, constructs a linear decision boundary, and then predicts the risk level of the transaction. Its prediction calculation relies on the linear combination relationship of features and parameters. However, the linear model cannot automatically capture the non-linear correlation between features, especially in the context of multi-dimensional and dynamic characteristics of fraudulent behavior, the static linear feature combination is difficult to cover complex risk patterns.

[0004] The tree-based ensemble model can implicitly learn the interaction between features by constructing multiple decision trees and integrating the prediction results of each tree. Its core logic is to generate decision trees sequentially, so that each tree learns the prediction residual of the previous tree, and finally accumulates the outputs of all trees to obtain the risk prediction result. However, the feature interaction learning depth of this type of model is limited, making it difficult to capture high-order feature interaction relationships, and lacking effective integration ability for professional scene knowledge in the field of financial risk control, resulting in a lack of business orientation in feature interaction and an inability to dynamically adjust interaction strategies for different scenarios.

[0005] The two types of existing technical solutions also have the following common problems: first, feature engineering and model training are independent of each other, lacking a bidirectional feedback mechanism, feature construction relies on human experience, and the feature interaction method cannot be dynamically optimized according to the risk prediction performance of the model, making it difficult to improve the model's ability to identify complex risks; second, the model lacks adaptability and the ability to track the dynamic evolution of risk patterns. Financial fraud methods continue to evolve, but once the model training is complete, its structure and parameters are fixed, and it cannot be self-adapted to match new risk patterns, resulting in a significant decline in model performance over time; third, the complexity of feature interaction does not match the risk level, either missing key risk features due to insufficient interaction depth or causing high computational costs due to excessive complexity, making it difficult to balance model performance and computational efficiency.

[0006] In summary, the existing financial transaction risk prediction technology cannot effectively cope with the dynamics of the financial transaction scene, the evolution of the risk pattern, and the professionalism of the business needs. SUMMARY

[0007] At least one embodiment of the present application provides a financial transaction risk control method, device, storage medium and program product, which are used to solve the problem that the existing financial transaction risk prediction technology cannot effectively cope with the dynamics of the financial transaction scene, the evolution of the risk pattern, and the professionalism of the business needs.

[0008] To solve the above technical problems, the present application is implemented as follows:

[0009] In a first aspect, the embodiments of the present application provide a financial transaction risk control method, comprising:

[0010] obtaining original data of a financial transaction and performing data preprocessing to output a standardized feature vector;

[0011] determining a scene probability and an enhanced feature according to the standardized feature vector and a preset scene adaptive attention mechanism;

[0012] determining an interaction intensity coefficient according to the enhanced feature and the scene probability;

[0013] inputting the enhanced feature and the interaction intensity coefficient into a prediction model to obtain a risk prediction result of risk prediction;

[0014] performing multi-level decision-making according to the risk prediction result and the scene probability to determine a target output result corresponding to the original data.

[0015] Optionally, obtaining original data of a financial transaction and performing data preprocessing to output a standardized feature vector comprises:

[0016] obtaining original data of a financial transaction;

[0017] performing data cleaning, missing value processing, feature encoding and conversion, and feature standardization on the original data in sequence to output a standardized feature vector;

[0018] wherein the data cleaning adopts 3σ criterion and box plot to detect abnormal values, verify data consistency, and remove duplicate records; the missing value processing fills in based on preset filling methods corresponding to numerical features, categorical features, and time series features; the feature encoding and conversion includes category feature encoding, time feature extraction, and geographic feature processing; the feature standardization operation includes numerical feature standardization and basic feature derivation; and the standardized feature vector includes a basic feature vector, a derived feature vector, an auxiliary information vector, and feature metadata.

[0019] Optionally, according to the standardized feature vector and a preset scene adaptive attention mechanism, a scene probability and an enhanced feature are determined, comprising:

[0020] A multi-layer perception structure is adopted to perform scene classification on the standardized feature vector, and a scene probability of a scene to which the financial transaction belongs is outputted;

[0021] According to the scene probability, an activation strength of a scene on an interaction template is calculated through a preset scene adaptive attention mechanism; the scene adaptive attention mechanism at least includes a risk sensitivity adjustment factor corresponding to the scene, a time decay factor based on a historical use time interval of the interaction template, and a scene-template interaction history vector;

[0022] After weight distribution of an interaction feature generated by the interaction template, filtering of the interaction template whose activation strength exceeds a preset threshold, and control of interaction depth, an enhanced feature is generated.

[0023] Optionally, after the enhanced feature is generated, the method further comprises:

[0024] After the interaction feature is generated, a weight of the interaction feature is determined based on a real-time risk status of the financial transaction; the weight adjustment is combined with the activation strength and performance change of the interaction feature in the current batch transaction;

[0025] According to a variance corresponding to the scene probability, an optimization period is set; the greater the variance, the shorter the optimization period;

[0026] According to the optimization period and the weight, the enhanced feature is regenerated.

[0027] Optionally, according to the enhanced feature and the scene probability, an interaction intensity coefficient is determined, comprising:

[0028] According to the enhanced feature, a basic risk score of the transaction is calculated;

[0029] Based on the scene probability and a risk adjustment coefficient corresponding to each scene, the basic risk score is adjusted to determine a comprehensive risk assessment result; the risk adjustment coefficient is determined according to a fraud rate of each scene in historical data;

[0030] According to the comprehensive risk assessment result and a preset intensity amplification coefficient, the interaction intensity coefficient is determined.

[0031] Optionally, the enhanced feature and the interaction intensity coefficient are inputted into a prediction model to obtain a risk prediction result of risk prediction, comprising:

[0032] The enhanced feature vector and the interaction intensity coefficient are input into a prediction model based on a light gradient boosting machine to perform risk prediction, and a risk prediction result of risk prediction is obtained.

[0033] In the prediction model, a learning rate in a model parameter is amplified according to a benchmark interaction intensity coefficient, a feature sampling rate in the model parameter is increased as the interaction intensity coefficient increases, and a minimum sample number of a leaf node in the model parameter is reduced as the interaction intensity coefficient increases.

[0034] Optionally, according to the risk prediction result and the scene probability, a multi-level decision is performed to determine a target output result corresponding to the original data, including:

[0035] A dynamic decision threshold is determined based on a preset benchmark decision threshold, a sum of products of scene probabilities of each transaction scene and corresponding scene threshold adjustment amounts, and the sum of products is added.

[0036] The risk prediction result is compared with the dynamic decision threshold, and in a case where the risk prediction result is greater than a sum of the dynamic decision threshold and a preset value, the target output result corresponding to the original data is determined as a direct rejection of transaction, and a detailed log containing transaction basic information and risk factors is recorded.

[0037] In a case where the risk prediction result is greater than the dynamic decision threshold and less than or equal to a sum of the dynamic decision threshold and the preset value, the target output result is triggered for secondary verification, and the transaction is allowed after the verification is passed.

[0038] In a case where the risk prediction result is less than or equal to the dynamic decision threshold, the target output result is to allow the transaction to pass normally.

[0039] In a second aspect, the embodiments of the present application provide a financial transaction risk control device, including:

[0040] A first processing module is configured to obtain original data of a financial transaction and perform data preprocessing, and output a standardized feature vector.

[0041] A first determination module is configured to determine a scene probability and an enhanced feature according to the standardized feature vector and a preset scene adaptive attention mechanism.

[0042] A second determination module is configured to determine an interaction intensity coefficient according to the enhanced feature and the scene probability.

[0043] A second processing module is configured to input the enhanced feature and the interaction intensity coefficient into a prediction model to obtain a risk prediction result of risk prediction.

[0044] A third determining module is configured to perform multi-level decision-making according to the risk prediction result and the scene probability, and determine the target output result corresponding to the original data.

[0045] In a third aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the steps of the method in any one of the first aspect.

[0046] In a fourth aspect, an embodiment of the present application provides a computer program product, which includes computer instructions. The computer instructions are executed by a processor to implement the steps of the method in any one of the first aspect.

[0047] Compared with the prior art, the financial transaction risk control method, device, storage medium and program product provided by the embodiment of the present application can solve the defects of the prior art in scene dynamics, risk pattern evolution and business demand specialization through the full-link design of data preprocessing, scene perception, intensity adaptation, dynamic prediction and hierarchical decision-making. The dimensional differences of the original data are eliminated through standardization processing, and a standardized feature vector is output, which provides high-quality data support for subsequent scene adaptation and risk identification, and avoids scene adaptation deviation caused by data format confusion. Based on the preset scene adaptive attention mechanism, the scene-related features in the standardized feature vector are dynamically weighted, the traditional one-size-fits-all feature processing mode is broken, and the dynamic risk differences of different scenes are adapted. The enhanced features and the scene probability are combined to calculate the interaction intensity coefficient, so as to ensure that the model invests more in fine calculation for new fraud patterns, and realizes dynamic adaptation of risk pattern evolution. The enhanced features and the interaction intensity coefficient are input into the prediction model, the model parameters are dynamically adjusted with the coefficient, which not only ensures the risk identification accuracy, but also takes into account the system efficiency, and copes with the prediction challenges brought by the evolution of risk patterns. Based on the risk prediction result and the scene probability, a dynamic decision threshold is calculated, which not only matches the risk tolerance of different business scenes, but also reduces the false judgment of normal transactions through secondary verification, and meets the professional differences of business needs. BRIEF DESCRIPTION OF DRAWINGS

[0048] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not intended to be limiting in any respect. Moreover, the use of the same reference symbols in different drawings indicates similar or identical items.

[0049] Figure 1 A flowchart of a financial transaction risk control method provided by an embodiment of the present application is shown in FIG. 1.

[0050] Figure 2 A structural diagram of a financial transaction risk control system provided by an embodiment of the present application is shown in FIG. 2.

[0051] Figure 3 A structural schematic diagram of the financial transaction risk control device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0052] The terms "first", "second", and the like in the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second" are generally of a kind and do not limit the number of objects, for example, the first object can be one or more. In addition, "or" in the present application means at least one of the connected objects. For example, "A or B" covers three scenarios, namely, scenario one: including A and not including B; scenario two: including B and not including A; scenario three: including A and B. The character " / " generally represents that the objects before and after are in an "or" relationship.

[0053] The term "indication" in the present application can be a direct indication (or explicit indication) or an indirect indication (or implicit indication). The direct indication can be understood as the sender explicitly informing the receiver of specific information, operations to be performed or requested results, etc. in the sent indication. The indirect indication can be understood as the receiver determining the corresponding information according to the indication sent by the sender, or judging and determining the operation to be performed or the requested result according to the judgment result.

[0054] In order for those skilled in the art to better understand the embodiments of the present application, the following is first explained:

[0055] Traditional models usually adopt a static feature interaction mode, which cannot capture the dynamic complex relationship between features in financial transactions, especially in the context of changing fraud behaviors. Linear models cannot automatically discover feature interactions, while tree-based models can implicitly learn feature interactions, but have limited depth and are difficult to capture high-order interaction relationships.

[0056] Existing models are difficult to effectively integrate professional knowledge in the field of financial risk control, resulting in a lack of business scenario guidance for feature interaction. Financial transaction risks are often closely related to specific scenarios, such as large-scale transactions in different places, frequent small-scale transfers, etc., but existing models cannot dynamically adjust feature interaction strategies according to different scenarios.

[0057] In the face of changing risk patterns, existing technologies lack adaptive adjustment mechanisms, resulting in a decline in model performance over time. Financial fraud methods are constantly evolving, but once the model is trained, the structure and parameters are fixed, making it difficult to adapt to new risk patterns in a timely manner.

[0058] Traditional methods take feature engineering and model training as independent steps, lack effective feedback mechanism, and are difficult to optimize feature interaction. Feature engineering relies on manual experience and cannot automatically adjust feature construction strategies according to model performance.

[0059] The problems faced by the above cases are solved by the embodiments of the present application. The financial transaction risk control method, device, storage medium and program product provided by the embodiments of the present application can reduce or avoid the occurrence of the above cases.

[0060] The embodiments of the present application provide a financial transaction risk control method, device, storage medium and program product. The method and the device are based on the same application concept. Since the principles of solving problems are similar, the implementation of the device and the method can be mutually referred to, and the repeated parts will not be described again.

[0061] Please refer to Figure 1 The financial transaction risk control method provided by the embodiments of the present application comprises:

[0062] Step 11, obtaining the original data of the financial transaction and performing data preprocessing to output a standardized feature vector;

[0063] Step 12, determining a scene probability and an enhanced feature according to the standardized feature vector and a preset scene adaptive attention mechanism;

[0064] Step 13, determining an interaction intensity coefficient according to the enhanced feature and the scene probability;

[0065] Step 14, inputting the enhanced feature and the interaction intensity coefficient into a prediction model to obtain a risk prediction result of risk prediction;

[0066] Step 15, performing multi-level decision according to the risk prediction result and the scene probability to determine a target output result corresponding to the original data.

[0067] In the embodiments of the present application, through the whole-link design of data preparation in step 11, scene perception in step 12, feature association in step 13, risk prediction in step 14, and dynamic decision-making in step 15, the pain points of poor adaptability, rigid feature weight, and single decision-making in traditional risk control scenes are solved, and the risk identification precision and risk control flexibility are improved in combination with scene self-adaptation and multi-level decision-making. In step 11, the original financial transaction data is obtained and preprocessed, and a standardized feature vector is output, completing the data preparation of the risk control model, converting the disordered original transaction data into standardized features that can be input into the model, and laying a data foundation for subsequent risk analysis. Step 12 realizes scene perception and solves the problem of one-size-fits-all feature weight in traditional risk control. The risk patterns of different financial scenes are very different, and key risk features need to be extracted accordingly. Step 13 realizes the correlation and quantization of scenes and features, further refines the influence of features on risk prediction in different scenes, and avoids the rigid problem of the same feature affecting the same in all scenes.

[0068] It should be noted that the interaction intensity coefficient is used to represent the contribution of a certain enhanced feature to the risk result in the current scene. The larger the interaction intensity coefficient, the more core the feature is as a risk indicator in the current scene; the smaller the interaction intensity coefficient, the smaller the influence of the feature on the risk judgment of the current scene. When the subsequent model is input, the enhanced feature is multiplied by the interaction intensity coefficient, which is equivalent to dynamically adjusting the feature weight. For example: the same enhanced feature of a transaction amount of 100,000 yuan has a much greater risk contribution in the cross-border transfer scene (interaction intensity coefficient 0.8) than in the daily consumption scene (interaction intensity coefficient 0.2), which is more in line with the actual risk control logic.

[0069] Step 14 completes risk prediction calculation, outputs the risk level or risk probability of the transaction based on the scene-optimized features, and is the core judgment link of the risk control link. In step 14, the prediction model can be selected, and a machine learning or deep learning model suitable for financial risk control can be selected. The model needs to be pre-trained and optimized based on historical risk transaction data and normal transaction data. The model input is the weighted feature vector obtained by multiplying the enhanced feature vector and the corresponding interaction intensity coefficient, which is used as the model input. The model output of step 14 is the risk prediction result: the model output can be two types of results: risk probability and / or risk level. The risk probability is used to represent the probability that the transaction is a risk transaction; the risk level is used to represent the level based on the probability division.

[0070] Step 15 realizes dynamic risk control decision, solves the problem of single risk control of high-risk interception and low-risk passing, and balances risk prevention and control and user experience. Combined with the scene probability and risk prediction result of the current transaction, the above decision rules are matched to output the final risk control action. For example, if the scene probability of a transaction is 85% for daily consumption and 10% for cross-border transfer, and the risk probability is 85% (medium risk), the target output result can be selected as: triggering SMS verification, completing the transaction if the verification is passed, and intercepting if the verification fails. For example, if the scene probability of a transaction is 90% for credit card cashing and the risk probability is 95% (high risk), the target output result is to freeze the user's credit card account, send a risk warning SMS to the user, and notify the risk control personnel for manual verification.

[0071] The data preprocessing of step 11 of the present application is the basis, providing high-quality input for the follow-up; the scene perception of step 12 is the core optimization point, realizing the scene enhancement of features; the interaction intensity coefficient of step 13 is the bridge, quantifying the risk association of scene and features; the risk prediction of step 14 is the core judgment, outputting the risk result; the multi-level decision of step 15 is the landing action, realizing differentiated risk control, forming a closed loop from data to decision through the five steps, and taking into account the risk control accuracy and user experience.

[0072] Further, the above step 11 includes:

[0073] Obtaining raw data of financial transactions;

[0074] Performing data cleaning, missing value processing, feature encoding and conversion, and feature standardization operation on the raw data in sequence, and outputting a standardized feature vector;

[0075] Wherein, the data cleaning adopts 3σ criterion and box plot to detect abnormal values, verify data consistency and remove duplicate records; the missing value processing fills in the pre-set filling mode corresponding to numerical features, categorical features and time series features; the feature encoding and conversion includes category feature encoding, time feature extraction and geographic feature processing; the feature standardization operation includes numerical feature standardization and basic feature derivation; the standardized feature vector includes a basic feature vector, a derived feature vector, an auxiliary information vector and feature metadata.

[0076] Here, the data type of the obtained raw data of financial transactions includes but is not limited to:

[0077] Transaction basic data: including transaction amount, transaction timestamp, transaction type (transfer, payment, withdrawal, etc.), transaction channel (online banking, mobile banking, ATM, etc.), merchant code, transaction status and other core transaction information.

[0078] User behavior data: user account basic information, historical transaction statistical features (average daily transaction times, average monthly transaction amount, common transaction time period, etc.), user risk level label, account opening duration, real-name authentication status, etc.

[0079] Device environment data: device fingerprint information, IP address and geographic location, operating system type, browser version, network connection method, device login history, etc. Technical environment features.

[0080] External auxiliary data: blacklist library information, geographic location risk level, time dimension features (workday, holiday, business hours, non-business hours), exchange rate information (cross-border transaction), etc.

[0081] Further, the data preprocessing operation includes: (1) data cleaning and quality control: outlier detection: identify outliers in transaction amount, frequency and other numerical features based on 3σ criterion and box plot method; data consistency check: verify the reasonableness of transaction time, the uniformity of amount accuracy, and the validity of user ID; duplicate data processing: remove duplicate records based on transaction unique identifier, and keep the record with the latest timestamp. (2) Missing value processing strategy: numerical features: use median to fill in key features such as transaction amount; statistical features: use 0 to fill in; categorical features: use "unknown" category to fill in merchant information, etc.; user label uses the most frequent category to fill in; time series features: use a combination of forward filling and backward filling method. (3) Feature encoding and conversion: category feature encoding: use One-Hot encoding to process low base features such as transaction type and channel, and use Label encoding to process high base features such as merchant ID; time feature extraction: extract periodic features such as hour, week, month from timestamp, and construct time distance features; geographic feature processing: convert IP address to geographic location code, and calculate distance features of user's frequently used locations.

[0082] Further, feature standardization and derivation operations include: (1) numerical feature standardization: apply Min-Max normalization to transaction amount, statistical features, etc. ; Wherein, X represents the original feature value, i.e. the user behavior data to be processed, which is the input object of normalization; represents the minimum value in the sample set; represents the maximum value of the numerical feature in the sample set, and correspond to the same batch of sample sets, ensuring consistency of the range; represents the normalized standard feature value. Apply Z-score standardization to user behavior statistical features: ; Wherein, represents the mean value of the feature in the sample set; denotes the standard deviation of the feature in the sample set; outlier truncation: extreme values exceeding the 95th percentile are truncated to the 95th percentile value. (2) Basic feature derivation: time window statistical features: calculate the number of transactions, transaction amount, etc. of the user within 1 hour, 1 day, 7 days; ratio features: the ratio of the current transaction amount to the user's historical average amount, the ratio to the daily average amount, etc.; risk behavior features: out-of-town transaction identifier, large transaction identifier, frequent transaction identifier, etc. binary features.

[0083] Further, the output format specification includes: output standardized feature vector , wherein is the feature dimension (usually 200-300 dimensions), is the batch size. The output includes:

[0084] Basic feature vector: standardized original features, dimension about 100-150;

[0085] Derived feature vector: time window statistics and ratio features, dimension about 50-100;

[0086] Auxiliary information vector: scene identification, user risk level, etc., dimension about 20-50;

[0087] Feature metadata: feature name, data type, missing value identifier, etc., used for subsequent feature selection.

[0088] The preprocessed data of the application is directly input to the scene-aware interaction network, ensuring the consistency of feature quality and format, and providing a reliable data basis for subsequent adaptive feature interaction.

[0089] Further, the above step 12 includes:

[0090] Using a multi-layer perception structure, the standardized feature vector is classified according to the scene, and the scene probability of the financial transaction belonging to the scene is output;

[0091] According to the scene probability, the activation strength of the scene on the interaction template is calculated through a preset scene adaptive attention mechanism; the scene adaptive attention mechanism at least includes: a risk sensitivity adjustment factor based on the scene, a time decay factor based on the historical use time interval of the interaction template, and a scene-template interaction history vector;

[0092] Using the activation strength, the interaction features generated by the interaction template are weighted and allocated, the interaction templates with activation strength exceeding a preset threshold are filtered, and the interaction depth is controlled, and then the enhanced features are generated.

[0093] In the embodiment of the application, the scene recognition layer adopts a multi-layer perception structure to classify the original features according to the scene:

[0094] ; wherein, represents the original feature vector; represents the first class of scenarios; and respectively represent model parameters; represents the total number of predefined scene categories; the financial transaction is divided into 8 typical scenarios by scene recognition: large amount transfer, small amount high frequency transaction, cross-border transaction, first transaction of new user, out-of-town transaction, sensitive period transaction, merchant payment and investment and financial scene. Each scenario represents different risk patterns and feature interaction requirements.

[0095] For the identified scene , the interaction template activation uses an improved scene adaptive attention mechanism. Unlike traditional attention mechanisms, the present application introduces risk sensitivity adjustment and time decay factor:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] wherein, represents the first interaction template; represents the activation strength of the scene to the template ; is the risk sensitivity adjustment factor, is the scene risk level; is the time decay factor, is the historical use time interval of the template in the scene ; is the scene-template interaction history vector; is the network parameter; is the decay coefficient.

[0101] Further, the calculated activation strength is directly used to guide the generation process of feature interaction: interaction feature weight allocation: the activation strength as a weight coefficient, determines the contribution of each interaction template to the final feature vector : ; wherein, is the first interaction template.The interaction feature vector generated by each template. Dynamic feature selection: set the activation threshold Only templates with intensity exceeding the threshold are activated to participate in feature generation, optimizing the allocation of computing resources.

[0102] Interaction depth control: dynamically adjust the complexity of feature interaction according to the activation intensity: ; wherein, is the interaction depth of the template , and is the maximum depth increment.

[0103] Optionally, after generating the enhanced features, the method further comprises:

[0104] After generating the interaction features, determine the weight adjustment of the interaction features based on the real-time risk status of financial transactions; the weight adjustment combines the activation intensity and the performance change of the interaction features in the current batch of transactions;

[0105] According to the variance corresponding to the scene probability, set the optimization period; the greater the variance, the shorter the optimization period; according to the optimization period and the weight, regenerate the enhanced features.

[0106] In the embodiments of the present application, the feedback weight can be updated, and the activation intensity participates in the feedback optimization of the model: templates with high activation intensity but low effectiveness will be adjusted: , wherein, represents the weight of a certain feature interaction template at time t (before updating), which measures the contribution of the interaction template to the current risk prediction model; the greater the weight, the more the model relies on the output of the template when judging the risk of transactions; represents the gradient of the model loss function with respect to the weight of the interaction template, which measures the influence of the small change of the template weight on the model risk prediction loss; if the gradient is positive, it means that the current weight increases the model loss (the template is ineffective); if the gradient is negative, it means that the current weight reduces the model loss (the template is effective); represents the activation intensity of the feature interaction template; represents the weight of the feature interaction template at time t+1 (after updating).

[0107] Wherein, the interaction generation layer designs multiple interaction templates, covering time-amount interaction, geographic-behavior interaction, user-merchant interaction, device-environment interaction, etc. Through accurate control of the activation intensity, scene adaptation and computing efficiency optimization of feature interaction are realized. The activation intensity and the interaction feature received by the interaction generation layer are used to dynamically adjust the feature weight based on the real-time risk status of financial transactions. An adaptive mapping relationship between transaction risk and feature importance is established.

[0108] Feature weight updating mechanism based on activation intensity: ; wherein, represents the original weight of the i-jth interaction feature at t, measuring the initial importance of the feature in the model, such as the initial weight of "time-amount interaction"; represents the updated feature weight at t+1; represents the weight update learning rate; represents the activation intensity of the i-jth interaction feature, from the interaction generation layer, reflecting the degree of association of the feature with the current transaction scenario, the higher the activation intensity, the more the feature meets the current scenario demand; is the performance change of the interaction feature in the current batch of transactions, reflecting the dynamic evolution of risk patterns in the financial market. When the risk pattern of a certain type of transaction (such as night large amount transfer) changes, the related time-amount interaction feature weight will be adjusted accordingly. Scenario volatility-aware optimization period setting: ; wherein, is the variance of the scenario recognition probability, reflecting the uncertainty of the current transaction environment; represents the basic optimization period, the default optimization interval when the scenario is stable, determined by historical scenario volatility data; represents the scenario volatility coefficient, controlling the influence range of scenario variance on optimization period, avoiding the dramatic change of period caused by small amplitude change of volatility; represents the dynamic optimization period, the time interval for the model to perform feature weight updating and parameter adjustment, and the output result. In the period of rapid evolution of fraud means (such as the emergence of new network fraud), the scenario probability distribution fluctuates greatly, and the system automatically shortens the optimization period to speed up the model adaptation speed.

[0109] Optionally, step 13 above comprises:

[0110] According to the enhanced feature, calculate the basic risk score of the transaction;

[0111] Based on the scenario probability and the risk adjustment coefficient corresponding to each scenario, adjust the basic risk score to determine the comprehensive risk assessment result; the risk adjustment coefficient is determined according to the fraud rate of each scenario in the historical data;

[0112] According to the comprehensive risk assessment result and the preset intensity amplification coefficient, determine the interaction intensity coefficient.

[0113] In the embodiments of the present application, the financial risk comprehensive assessment is based on the enhanced feature vector to calculate the risk score of the transaction : , wherein, represents the basic risk score of the transaction x; represents the activation function; represents an enhanced feature weight matrix; represents an enhanced feature vector; represents a bias term.

[0114] Risk adjustment is combined with scenario probability: ; represents the final risk score of transaction x; represents the risk adjustment coefficient of the i-th scenario; represents the sum of scenario risk adjustments; represents the probability that transaction x belongs to the i-th scenario. Wherein, Risk adjustment coefficients for each scenario: large transfer scenario , cross-border transaction scenario , first transaction of new user , etc. These coefficients are determined according to the fraud rate of each scenario in historical data.

[0115] Dynamic adjustment of computing resource allocation based on risk assessment results: , wherein, represents the interaction intensity coefficient of transaction x, which controls the degree of resource investment for subsequent feature interaction and model calculation. The larger the coefficient, the more refined the calculation (such as deeper LightGBM tree and higher feature sampling rate); represents an upper bound constraint function to ensure that the interaction intensity coefficient does not exceed 1.0. Even if >1.0, only 1.0 resources are allocated to avoid high-risk transactions occupying too much computing resources causing system delay; is the intensity amplification coefficient to ensure that high-risk transactions receive more refined feature interaction analysis. This intensity coefficient directly affects the feature selection and tree construction depth of the subsequent LightGBM model.

[0116] Optionally, for high-risk transactions ( ): use the complete set of interaction features, and set the maximum depth of LightGBM to 8 layers; for medium-risk transactions ( ): select the top-60% important interaction features, and set the maximum depth to 6 layers; for low-risk transactions ( ): only use core interaction features, and set the maximum depth to 4 layers.

[0117] Optionally, input the enhanced features and the interaction intensity coefficient into the prediction model to obtain the risk prediction result of risk prediction, including:

[0118] Input the enhanced feature vector and the interaction intensity coefficient into the prediction model based on the lightweight gradient boosting machine to perform risk prediction and obtain the risk prediction result of risk prediction;

[0119] Among them, in the prediction model, the learning rate in the model parameter is amplified according to the interaction intensity coefficient of the benchmark, the feature sampling rate in the model parameter is improved with the increase of the interaction intensity coefficient, and the minimum sample number of the leaf node in the model parameter is reduced with the increase of the interaction intensity coefficient.

[0120] In the embodiments of the application, the enhanced feature vector (containing basic features and interaction features) and the interaction intensity coefficient are input into a LightGBM model in linkage, model parameters are dynamically adjusted to adapt to transactions of different risk levels, and finally accurate risk prediction results are output. High-risk transactions are identified using more refined model parameters, low-risk transactions are processed quickly using simplified parameters, and the risk identification accuracy and system efficiency are balanced.

[0121] Specifically, the input data is determined: the enhanced feature vector and the interaction intensity coefficient. The input enhanced feature vector refers to a comprehensive feature set after preprocessing (such as normalization and standardization) and feature interaction (such as time-amount, geographic-behavior interaction), which contains the basic attributes of transactions (amount, time, device), user historical behavior (transaction frequency in the past 30 days), scene-related features (whether cross-border, whether large amount), etc., and is the core basis for model risk judgment. The interaction intensity coefficient also needs to be input, which is calculated by the formula in the foregoing, and reflects the degree of model calculation resource investment required for the current transaction.

[0122] A prediction model is selected, and LightGBM is selected in the present application. LightGBM is an integrated learning model commonly used in the field of financial risk control, has the advantages of fast training speed, strong processing capability for high-dimensional features, good interpretability (through feature importance), and is suitable for processing real-time risk prediction of millions of transaction data. Through integrated learning of multiple decision trees, the enhanced feature vector is mapped to a risk prediction result (i.e., the probability of a transaction being fraudulent) in the 0-1 interval.

[0123] The adjustment logic of the model parameters is that the larger the interaction intensity coefficient (the higher the risk), the more complex the model (the more refined the identification), and the specific parameter adjustment rules are as follows:

[0124] The learning rate is adjusted based on the benchmark learning rate (such as 0.01) under the condition that the adjustment rule increases with the increase of the interaction intensity coefficient (the larger the learning rate, the higher the learning rate). High-risk transactions require the model to converge to a fine feature mode faster, improving identification sensitivity.

[0125] The feature sampling rate is improved under the condition that the adjustment rule increases with the increase of the interaction intensity coefficient (when λ=1.0, the sampling rate=0.9, and when λ=0.3, the sampling rate=0.5). High-risk transactions require the use of more features (such as full-amount interaction features) to avoid missing key risk signals; low-risk transactions can use core features, saving calculation.

[0126] The minimum number of leaf nodes is reduced (λ=1.0, =10; λ=0.3, =50) as the interaction intensity coefficient increases. High-risk transactions allow the decision tree to grow deeper (smaller leaf nodes) to capture more fine-grained risk patterns (such as fraud at 3 a.m. and specific virtual merchants). Low-risk transactions limit the tree depth to avoid overfitting and time-consuming calculations. The model calculates the input enhanced feature vector using the dynamically adjusted parameters and outputs the risk prediction result (e.g., risk prediction result =0.75, which means there is a 75% probability that the transaction is fraudulent).

[0127] The application accurately identifies high-risk transactions: high-risk transactions can capture subtle fraud features through complex model parameters (high sampling rate, deep tree structure), improving recognition accuracy; improve low-risk transaction processing efficiency: low-risk transactions reduce computation time through simplified parameters, ensuring that the system can handle millions of daily transactions without delay; parameter adaptation: no manual adjustment of model parameters, the system automatically adapts according to transaction risk levels, reducing operational costs.

[0128] Optionally, according to the risk prediction result and the scene probability, a multi-level decision is performed to determine the target output result corresponding to the original data, comprising:

[0129] Based on the preset baseline decision threshold, the sum of the product of the scene probability of each transaction scene and the corresponding scene threshold adjustment amount is determined to determine the dynamic decision threshold;

[0130] Compare the risk prediction result with the dynamic decision threshold. If the risk prediction result is greater than the sum of the dynamic decision threshold and a preset value, determine that the target output result corresponding to the original data is to directly reject the transaction, and record detailed logs containing transaction basic information and risk factors;

[0131] If the risk prediction result is greater than the dynamic decision threshold and less than or equal to the sum of the dynamic decision threshold and the preset value, the target output result is to trigger secondary verification, and the transaction is allowed after verification;

[0132] If the risk prediction result is less than or equal to the dynamic decision threshold, the target output result is to allow the transaction to pass normally.

[0133] In the embodiment of the application, the risk prediction result and the scene probability are combined, the dynamic decision threshold is used instead of the fixed threshold, different scene differentiated risk control decisions are realized, and finally three target results of direct rejection, secondary verification, and allowed to pass are output, balancing risk interception and user experience (such as avoiding normal large amount of money being misintercepted).

[0134] Specifically, a dynamic decision threshold is calculated. The dynamic decision threshold is the baseline threshold and the scene adjustment amount, and the core logic is that the threshold is lowered in a high-risk scenario (easier to intercept), and the threshold is raised in a low-risk scenario (more relaxed), based on the predicted probability and the scene probability The dynamic decision threshold is set, and the calculation formula is as follows: ; wherein, is the threshold adjustment amount of each scene: the threshold is lowered by 0.05 in a sensitive period transaction, and the threshold is raised by 0.1 in a large amount transfer, etc., reflecting the risk tolerance difference of different transaction scenarios; represents the dynamic decision threshold of the xth transaction, which is used to judge whether the transaction is risky, and the predicted probability P_fraud(x) exceeds the threshold, which needs to trigger the risk control action; represents the basic decision threshold, which is the default baseline for all scenarios, and is obtained by training historical risk control data to balance the false rejection rate and the false acceptance rate; represents the probability that the xth transaction belongs to the ith scene .

[0135] The risk prediction result is compared with the dynamic decision threshold, and three types of decisions are executed according to different intervals, and the decision rules include: : direct rejection, operation: intercept the transaction, record detailed logs (including transaction amount, device information, risk features such as new user + out-of-place), and synchronize to the risk control background for manual review; for high-risk transactions (such as fraud probability much higher than the threshold), avoid financial loss; : trigger secondary verification (SMS or face recognition), require the user to complete additional verification (such as SMS verification code, face recognition, transaction password secondary input), and pass the verification to release, and fail to intercept; for medium-risk transactions (there is a certain fraud possibility but not absolute), exclude normal users (such as out-of-place large amount transfer operated by the user himself) through secondary verification. : normal pass, allow the transaction to proceed normally, directly release the transaction, and only record basic information (do not interfere with the user process); for low-risk transactions (fraud probability is low), ensure user experience (such as daily small amount consumption).

[0136] According to the above decision conditions, the system outputs the final target result (reject, verify or pass), and triggers the corresponding business process (such as intercepting the transaction, sending the verification SMS, and completing the payment).

[0137] In the embodiments of the present application, the precise prediction of the model is used for intelligent decision-making linkage. The first step is to output reliable risk prediction results through the model of dynamic parameters, the second step is to convert the prediction results into implementable risk control actions based on the scene dynamic threshold, and finally the financial risk control goal of accurate risk identification, reasonable decision-making and good user experience is realized.

[0138] ReferenceFigure 2 As shown, the embodiments of the present application also provide a financial transaction risk control system, which is based on the design concept of "parameter transmission chain and closed-loop feedback" and consists of five closely coupled core modules: a data preprocessing module, which is responsible for data cleaning, missing value processing and basic feature extraction, and outputs a standardized feature vector ; a scene-aware feature interaction module, which is a core innovative module and contains a scene recognition layer, an interaction generation layer and an adaptive optimization layer, and outputs a scene probability , an activation intensity and an enhanced feature ; a risk assessment and interaction intensity adjustment module, which calculates a comprehensive risk assessment based on the scene probability and the enhanced feature , and generates an interaction intensity coefficient to guide subsequent calculation resource allocation; a LightGBM prediction and feedback optimization module, which receives the enhanced feature and the intensity parameter for adaptive risk prediction , and generates a feedback signal to update the feature interaction; a decision mechanism and business application module, which makes multi-level decisions based on the prediction results and scene information, and outputs a business understandable risk explanation.

[0139] Further, the data preprocessing module is a basic component of the system, which is responsible for converting raw financial transaction data into a standardized feature vector suitable for subsequent scene-aware interaction network processing. This module mainly processes four categories of data sources.

[0140] Further, the scene-aware feature interaction module is the core innovation of the present application, which consists of three levels: a scene recognition layer, an interaction generation layer and an adaptive optimization layer.

[0141] Further, the risk assessment and interaction intensity adjustment module receives the enhanced feature vector generated by the scene-aware interaction network , performs comprehensive assessment based on the risk features of the financial transaction, and dynamically adjusts the calculation intensity of subsequent processing. Among them, refers to the core basic features selected from the standardized feature vector after step 11 preprocessing, such as transaction amount (after standardization), user historical risk score, transaction channel type (after coding), device trust level, etc. This kind of feature is the basic anchor point of risk assessment, reflects the core attributes of transaction, and has low calculation cost and strong interpretability. (Feature interaction derived feature) refers to cross-dimensional correlation features mined through scene adaptive attention mechanism, which is a risk information enhancement of .

[0142] The LightGBM prediction and feedback optimization module is used to perform adaptive LightGBM training and prediction result feedback optimization. Optionally, the adaptive LightGBM training includes: the LightGBM model receives the enhanced feature vector and the interaction intensity parameter to perform risk prediction: ; wherein, represents the interaction intensity coefficient; represents the enhanced feature vector; the model parameters are dynamically adjusted according to the interaction intensity: the learning rate ; the feature sampling rate ; the minimum number of leaf nodes .

[0143] Optionally, the prediction result feedback optimization includes: the system calculates a feedback signal for optimizing feature interaction: ; wherein, N represents the number of batch transaction samples participating in feedback calculation; n represents the index of a single transaction sample; represents the true risk label of the nth transaction; represents the model predicted fraud probability of the nth transaction; represents the i-jth feature interaction term of the nth transaction; represents the partial derivative of the predicted fraud probability with respect to the i-jth interaction feature, which is used to measure the influence sensitivity of the interaction feature on the prediction result. The larger the derivative, the stronger the influence of the feature change on the prediction probability.

[0144] The feedback signal directly updates the activation intensity of the interaction generation layer: ; forming a complete closed loop of scene recognition, feature interaction, risk prediction, and feedback optimization, ensuring that the system can continuously adapt to the evolution of financial fraud means. Among them, represents the updated activation intensity at the t+1th iteration, the result of feedback optimization, which serves as the basis for the next round of feature interaction; represents the activation intensity of the i-jth interaction feature at the tth iteration, which controls the importance of the interaction feature in the feature interaction layer. The greater the intensity, the more likely the feature is to be included in subsequent risk prediction; represents the feedback learning rate, which is used to control the update amplitude of the activation intensity to avoid system shock caused by excessive single update.

[0145] The decision mechanism and business application module is used to perform multi-level decision threshold setting and explainable output. Among them, the multi-level decision threshold setting includes: based on the prediction probability and the scene probability to set a dynamic decision threshold: ; wherein, Threshold adjustment amount for each scenario: sensitive period transaction reduction threshold 0.05, large amount transfer increase threshold 0.1, etc., reflecting the risk tolerance difference of different transaction scenarios; Dynamic decision threshold of the xth transaction, used to determine whether the transaction is risky, and the prediction probability P_fraud(x) exceeds the threshold, triggering the risk control action; The base decision threshold is the default benchmark for all scenarios, trained from historical risk control data, balancing the false negative rate and the false positive rate. The probability that the xth transaction belongs to the ith scenario.

[0146] Optionally, the decision rule includes: Directly reject and record detailed logs; Trigger secondary verification (SMS or facial recognition); Normally pass.

[0147] Optionally, the explainability output generates explanation information for each decision, including: main risk factors: Top-3 interaction features based on value sorting and their business implications; scenario analysis: dominant scenario type and its historical risk statistics; comparison benchmark: comparison with historical behavior of the user and the same user group.

[0148] For example: this transfer is identified as high risk, mainly because of the "night period x large amount" interaction anomaly (contribution 32%), "remote login x new payee" combination suspicious (contribution 28%), and "transaction frequency x amount change" pattern inconsistent (contribution 19%). This transaction belongs to the remote large amount transfer scenario, and the historical fraud rate of this scenario is 15.3%.

[0149] Specifically, the present application provides a complete parameter transmission chain: the system establishes an end-to-end parameter transmission mechanism to ensure the organic cooperation between modules:

[0150] ;

[0151] The output of each link is used as the key input of the next link, forming the logical main line of the technical solution.

[0152] The system of the present application establishes a complete technical closed loop of scenario recognition, feature interaction, risk prediction, and feedback optimization: forward propagation: the original data is gradually enhanced through each module to finally generate a risk prediction result; backward optimization: the prediction error is transmitted back to the feature interaction layer through feedback signals, updating the activation strength ; adaptive adjustment: dynamically adjusting the optimization frequency based on the volatility of the scenario, adapting to changes in financial risk patterns; ​​​​

[0153] The present application can be designed for multiple typical financial transaction scenarios (large amount transfer, cross-border transaction, first transaction of new user, etc.) with special processing logic, and the scene risk adjustment coefficient Based on historical fraud rate determination. Risk adaptive resource allocation: according to the risk level of the transaction Intelligent allocation of computing resources: high-risk transactions ( ): complete feature set, 8 layers deep, fine-grained analysis; medium-risk transactions ( ): 60% feature set, 6 layers deep, balanced efficiency; low-risk transactions ( ): core features, 4 layers deep, fast processing. Each decision provides risk factor explanation based on activation strength to meet the needs of financial regulation and business.

[0154] The present application provides prediction probability, scene information, decision mechanism, risk decision, and interpretable output. Through this tightly coupled module design and parameter passing mechanism, the system realizes the deep integration of technical innovation and financial business, ensuring the logical coherence and practical operability from technical implementation to business application.

[0155] The present application provides an improved scene adaptive attention mechanism. Based on the traditional attention mechanism, the technical scheme introduces risk sensitivity adjustment factor , time decay factor and scene-template interaction history vector to realize accurate perception of financial transaction scenarios and template activation.

[0156] The present application provides a multi-dimensional application mechanism for activation strength. Activation strength has four-dimensional specific application methods: interaction feature weight distribution: ; dynamic feature selection: threshold based computing resource optimization; interaction depth control: ; feedback weight update: . A complete technical link from calculation to application of activation strength is established, solving the problem of attention mechanism output without practical effect.

[0157] The present application dynamically adjusts the technical scheme of subsequent processing parameters according to the comprehensive risk assessment results , including: interaction intensity coefficient: ; LightGBM parameter adaptation: dynamic adjustment of learning rate, feature sampling rate, and minimum number of leaf nodes; computing resource hierarchical allocation: differentiated processing strategy for high / medium / low risk transactions. Realize the automatic mapping from risk assessment to model parameters, and intelligently allocate computing resources according to the risk level of financial transactions.

[0158] The technical solutions of the present application are based on the scenario recognition probability variance Dynamic adjustment of optimization period , and the feature weight updating mechanism based on activation intensity . The uncertainty of fraud evolution in the financial market is quantified, and the intelligent adjustment of model optimization frequency is realized.

[0159] The complete feedback loop of the present application from prediction results to feature interaction optimization: prediction error Feedback signal calculation Activation intensity update Feature interaction adjustment, forming a technical closed loop of scenario recognition, feature interaction, risk prediction, and feedback optimization. The self-learning and self-adaptive mechanism of the financial risk control model is established, ensuring that the system can continuously adapt to new fraud methods.

[0160] In summary, compared with existing models such as LightGBM and XGBoost, which use a unified feature interaction method and cannot distinguish the characteristics of different financial transaction scenarios, feature engineering and model training are independent. The present application proposes a scenario adaptive attention mechanism that directly integrates financial field knowledge (scenario risk level, historical usage patterns) into the feature interaction process, achieving deep integration of technical solutions and business scenarios. Risk sensitivity adjustment factor and time decay factor are introduced, enabling the model to automatically identify and adapt to the optimal feature interaction strategy in different risk scenarios.

[0161] Compared with the relative independence of each module in existing technologies, there is a lack of effective parameter transmission and collaborative optimization mechanism, resulting in limited overall performance improvement of the model. The present application establishes a complete parameter transmission chain: scenario probability , risk adjustment coefficient comprehensive risk assessment interaction intensity LightGBM parameter adjustment, the output of each link is used as the key input of the next link, forming an organic and unified technical system.

[0162] Compared with existing models that use the same computational complexity for all transactions, it not only wastes computing resources but also cannot provide sufficient detailed analysis for high-risk transactions. The present application intelligently allocates computing resources based on transaction risk level, using a complete feature set and 8 layers of depth for high-risk transactions ( ), and only using core features and 4 layers of depth for low-risk transactions ( ), improving processing efficiency by 20% while ensuring detection rate.

[0163] Compared with the prior art, the feature interaction mode is not in line with the actual business rules due to the lack of deep understanding of the financial transaction business scene. The business characteristics of 8 typical financial transaction scenes (large amount transfer, cross-border transaction, first transaction of new user, etc.) are directly encoded into the technical solution in the application, the scene risk adjustment coefficient is determined based on the historical fraud rate (0.8 for the first transaction of new user, 0.5 for cross-border transaction), and the decision threshold is adjusted according to the business risk tolerance, so as to realize the seamless combination of technology and business.

[0164] Compared with the prior art, once the model is trained, the structure and parameters are fixed, and it cannot automatically adapt to the rapid evolution of financial fraud means. The application constructs a complete closed-loop feedback mechanism, which automatically adjusts and optimizes the frequency through scene volatility perception The feature interaction is continuously optimized to ensure that the model can respond to new fraud patterns in a timely manner. Experimental results show that the new risk identification capability is greatly improved.

[0165] Compared with the prior art, either the explainability is strong but the performance is limited (such as logistic regression), or the performance is excellent but lacks explanation (such as deep learning), which is difficult to meet the requirements of financial supervision. The application ensures high performance while providing specific risk factor explanations based on the activation strength

[0166] The application has obvious commercial value, mainly reflected in the following aspects:

[0167] (1) Reduce financial losses. Through more accurate risk identification and lower false positive rate, financial institutions can significantly reduce direct economic losses caused by fraud. Experimental data show that the application can identify 65% of new fraud patterns in advance, which is much higher than the 23% of traditional models, which means that a large number of fraudulent transactions can be blocked in advance.

[0168] (2) Improve customer experience. By reducing false positives, the application significantly reduces interference to normal users, improves user satisfaction and retention rate. Financial institutions can provide a smoother transaction experience while ensuring security.

[0169] (3) Reduce operating costs. The application has high automation, which can reduce manual review links and reduce labor costs. At the same time, through the optimized allocation of computing resources, the system processing efficiency is improved, and the cost of computing resources is reduced.

[0170] The above introduces various methods of the embodiments of the application. The following will further provide a device for implementing the above method.

[0171] Please refer to Figure 3 The embodiments of the application also provide a financial transaction risk control device, which comprises:

[0172] ​​The first processing module 31 is configured to acquire original data of a financial transaction and perform data preprocessing, and output a standardized feature vector.

[0173] The first determination module 32 is configured to determine a scene probability and an enhanced feature according to the standardized feature vector and a preset scene adaptive attention mechanism.

[0174] The second determination module 33 is configured to determine an interaction intensity coefficient according to the enhanced feature and the scene probability.

[0175] The second processing module 34 is configured to input the enhanced feature and the interaction intensity coefficient into a prediction model, and acquire a risk prediction result of risk prediction.

[0176] The third determination module 35 is configured to perform multi-level decision-making according to the risk prediction result and the scene probability, and determine a target output result corresponding to the original data.

[0177] Optionally, the first processing module 31 comprises:

[0178] The first acquisition unit is configured to acquire original data of a financial transaction.

[0179] The first processing unit is configured to sequentially perform data cleaning, missing value processing, feature encoding and conversion, and feature standardization operation on the original data, and output a standardized feature vector.

[0180] In the data cleaning, 3σ criterion and box plot are used to detect abnormal values, verify data consistency, and remove duplicate records. In the missing value processing, preset filling methods corresponding to numerical features, categorical features, and time series features are used for filling processing. The feature encoding and conversion include category feature encoding, time feature extraction, and geographic feature processing. The feature standardization operation includes numerical feature standardization and basic feature derivation. The standardized feature vector includes a basic feature vector, a derived feature vector, an auxiliary information vector, and feature metadata.

[0181] Optionally, the first determination module 32 comprises:

[0182] The second processing unit is configured to use a multi-layer perception structure to perform scene classification on the standardized feature vector, and output a scene probability of a scene to which the financial transaction belongs.

[0183] The third processing unit is configured to calculate the activation strength of a scene on an interaction template according to the scene probability and through a preset scene adaptive attention mechanism. The scene adaptive attention mechanism at least includes a risk sensitivity adjustment factor corresponding to the scene, a time decay factor based on a historical use time interval of the interaction template, and a scene-template interaction history vector.

[0184] The fourth processing unit is configured to perform weight distribution on the interaction features generated by the interaction templates according to the activation intensity, filter the interaction templates whose activation intensity exceeds a preset threshold, and generate enhanced features after processing of the interaction depth.

[0185] Optionally, the device further comprises:

[0186] The fourth determination module is configured to determine adjustment of the weight of the interaction features based on real-time risk status of the financial transaction after the interaction features are generated; and the weight adjustment is combined with the activation intensity and performance change of the interaction features in the current batch transaction.

[0187] The third processing module is configured to set an optimization period according to the variance corresponding to the scene probability; the greater the variance, the shorter the optimization period.

[0188] The fourth processing module is configured to regenerate the enhanced features according to the optimization period and the weight.

[0189] Optionally, the second determination module 33 comprises:

[0190] The fifth processing unit is configured to calculate a basic risk score of the transaction according to the enhanced features.

[0191] The first determination unit is configured to adjust the basic risk score based on the scene probability and a risk adjustment coefficient corresponding to each scene to determine a comprehensive risk assessment result; the risk adjustment coefficient is determined according to a fraud rate of each scene in historical data.

[0192] The second determination unit is configured to determine an interaction intensity coefficient according to the comprehensive risk assessment result and a preset intensity amplification coefficient.

[0193] Optionally, the second processing module 34 comprises:

[0194] The sixth processing unit is configured to input the enhanced feature vector and the interaction intensity coefficient into a prediction model based on a light gradient boosting machine to perform risk prediction and obtain a risk prediction result of the risk prediction.

[0195] In the prediction model, a learning rate in a model parameter is amplified and adjusted according to the benchmark interaction intensity coefficient, a feature sampling rate in the model parameter is increased with the increase of the interaction intensity coefficient, and a minimum sample number of a leaf node in the model parameter is reduced with the increase of the interaction intensity coefficient.

[0196] Optionally, the third determination module 35 comprises:

[0197] The third determining unit is configured to determine a dynamic decision threshold by superimposing a sum of products of the scenario probabilities of each transaction scenario and the corresponding scenario threshold adjustment amount based on a preset reference decision threshold;

[0198] The seventh processing unit is configured to compare the risk prediction result with the dynamic decision threshold, and determine that the target output result corresponding to the original data is direct rejection of transaction and record a detailed log containing transaction basic information and risk factors in a case that the risk prediction result is greater than a sum of the dynamic decision threshold and a preset value.

[0199] The eighth processing unit is configured to determine that the target output result is triggering secondary verification and allowing transaction after verification in a case that the risk prediction result is greater than the dynamic decision threshold and less than or equal to a sum of the dynamic decision threshold and the preset value.

[0200] The ninth processing unit is configured to determine that the target output result is allowing transaction in a case that the risk prediction result is less than or equal to the dynamic decision threshold.

[0201] It should be noted that the device in the embodiment corresponds to the financial transaction risk control method described above, and the implementation manners in the above embodiments are applicable to the embodiment of the device and can achieve the same technical effects. The device provided in the embodiment can implement all method steps achieved by the method embodiments and can achieve the same technical effects. Therefore, the same parts and beneficial effects in the embodiment and the method embodiments will not be described in detail.

[0202] The embodiment of the present application further provides a computer readable storage medium, and the computer readable storage medium stores a computer program. The computer program is executed by a processor to implement various processes of the financial transaction risk control method embodiments and can achieve the same technical effects. To avoid repetition, the same parts and beneficial effects will not be described in detail. The computer readable storage medium includes a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0203] The embodiment of the present application further provides a computer program product including computer instructions. The computer instructions are executed by a processor to implement various processes of the financial transaction risk control method embodiments and can achieve the same technical effects. To avoid repetition, the same parts and beneficial effects will not be described in detail.

[0204] It should be noted that in the technical solutions of the present disclosure, the collection, collection, updating, analysis, processing, use, transmission, storage and the like of user personal information are in line with the provisions of relevant laws and regulations, are used for legal purposes, and do not violate public order and good customs. Necessary measures are taken for personal information to prevent illegal access to user personal information data and maintain user personal information security and network security.

[0205] It should be noted that in this paper, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive containing, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or device. Without more limitations, the element defined by the sentence "includes a" does not exclude the presence of another identical element in the process, method, article or device including the element.

[0206] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment method can be realized by software and necessary general hardware platform, of course, it can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for making a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) execute the method described in each embodiment of the present application.

[0207] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative, not limiting, and those skilled in the art can make many forms under the inspiration of the present application without departing from the scope of the present application and the scope of protection of the claims.

Claims

1. A financial transaction risk control method, characterized in that, include: Obtain the raw data of financial transactions and perform data preprocessing to output standardized feature vectors; Based on the standardized feature vector and the preset scene adaptive attention mechanism, the scene probability and enhanced features are determined; The interaction strength coefficient is determined based on the enhanced features and the scene probability; The enhanced features and the interaction strength coefficients are input into the prediction model to obtain the risk prediction result. Based on the risk prediction results and the scenario probabilities, multi-level decision-making is performed to determine the target output result corresponding to the original data.

2. The method according to claim 1, characterized in that, Obtain the raw data of financial transactions and perform data preprocessing to output standardized feature vectors, including: Obtain raw data from financial transactions; The original data is processed sequentially with data cleaning, missing value handling, feature encoding and transformation, and feature standardization operations, and a standardized feature vector is output. The data cleaning process employs the 3σ criterion and box plots to detect outliers, verify data consistency, and remove duplicate records. Missing value handling utilizes preset filling methods corresponding to numerical, categorical, and time-series features. Feature encoding and transformation include categorical feature encoding, time feature extraction, and geographic feature processing. Feature standardization operations include numerical feature standardization and basic feature derivation. The standardized feature vector includes basic feature vectors, derived feature vectors, auxiliary information vectors, and feature metadata.

3. The method according to claim 1, characterized in that, Based on the standardized feature vector and the preset scene adaptive attention mechanism, the scene probability and enhanced features are determined, including: A multilayer perceptron structure is used to classify the standardized feature vectors into scenarios and output the scenario probability of the financial transaction. Based on the scenario probability, the activation intensity of the scenario to the interaction template is calculated through a preset scenario adaptive attention mechanism; the scenario adaptive attention mechanism includes at least: a risk sensitivity adjustment factor based on the scenario, a time decay factor based on the historical usage time interval of the interaction template, and a scenario-template interaction history vector. Using the activation intensity, the interaction features generated by the interaction template are weighted, interaction templates with activation intensity exceeding a preset threshold are filtered, and the interaction depth is controlled to generate enhanced features.

4. The method according to claim 3, characterized in that, After generating the enhanced features, the method further includes: After generating the interaction features, the weights of the interaction features are determined based on the real-time risk status of financial transactions; the weight adjustment combines the activation intensity and the performance changes of the interaction features in the current batch of transactions. The optimization period is set based on the variance corresponding to the scenario probability; the larger the variance, the shorter the optimization period. The enhanced features are regenerated based on the optimization period and the weights.

5. The method according to claim 1, characterized in that, Based on the enhanced features and the scene probability, the interaction strength coefficient is determined, including: Calculate the underlying risk score of the transaction based on the enhanced features; Based on the scenario probabilities and the corresponding risk adjustment coefficients for each scenario, the basic risk score is adjusted to determine the comprehensive risk assessment result; the risk adjustment coefficients are determined based on the fraud rates of each scenario in historical data. The interaction intensity coefficient is determined based on the comprehensive risk assessment results and the preset intensity amplification coefficient.

6. The method according to claim 1, characterized in that, The enhanced features and the interaction strength coefficient are input into the prediction model to obtain the risk prediction result, including: The enhanced feature vector and interaction strength coefficient are input into the prediction model based on the lightweight gradient booster to perform risk prediction and obtain the risk prediction result. In the prediction model, the learning rate in the model parameters is amplified and adjusted according to the benchmark interaction strength coefficient, the feature sampling rate in the model parameters increases as the interaction strength coefficient increases, and the minimum number of samples for the leaf nodes in the model parameters decreases as the interaction strength coefficient increases.

7. The method according to claim 1, characterized in that, Based on the risk prediction results and the scenario probabilities, multi-level decision-making is performed to determine the target output results corresponding to the original data, including: Based on the preset baseline decision threshold, the dynamic decision threshold is determined by summing the product of the scenario probability of each transaction scenario and the corresponding scenario threshold adjustment amount. The risk prediction result is compared with the dynamic decision threshold. If the risk prediction result is greater than the sum of the dynamic decision threshold and the preset value, the target output result corresponding to the original data is determined to be a direct rejection of the transaction, and a detailed log containing basic transaction information and risk factors is recorded. If the risk prediction result is greater than the dynamic decision threshold and less than or equal to the sum of the dynamic decision threshold and the preset value, then the target output result is to trigger secondary verification, and the transaction is allowed after the verification is passed. If the risk prediction result is less than or equal to the dynamic decision threshold, the target output result is to allow the transaction to proceed normally.

8. A financial transaction risk control device, characterized in that, include: The first processing module is used to acquire the raw data of financial transactions, perform data preprocessing, and output standardized feature vectors. The first determining module is used to determine the scene probability and enhanced features based on the standardized feature vector and the preset scene adaptive attention mechanism; The second determining module is used to determine the interaction intensity coefficient based on the enhanced features and the scene probability; The second processing module is used to input the enhanced features and the interaction strength coefficient into the prediction model to obtain the risk prediction result of the risk prediction. The third determining module is used to perform multi-level decision-making based on the risk prediction results and the scenario probabilities to determine the target output results corresponding to the original data.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 7.

Images